Use a `~/.scarfbench` directory to store and retrieve the dataset as a sensible default

[rahlk]

Steps to follow--

1. Define .scarfbench layout (benchmark, evals, logs, tmp).
2. Add one shared path-resolution policy: flag > env > .scarfbench default.
3. Add bootstrap logic to create/validate .scarfbench directories.
4. Make bench pull default to .scarfbench/benchmark.
5. Make bench list and bench test default to .scarfbench/benchmark.
6. Make eval run default eval-out to .scarfbench/evals and benchmark input to .scarfbench/benchmark.
7. Keep existing flags working as explicit overrides (backward compatible).
8. Standardize path normalization/error handling across all commands.
9. Add tests for default-path behavior and override behavior.
10. Update CLI help + README examples to document new defaults.